September 1928. Alexander Fleming returns from holiday to find his laboratory in disarray. On the bench, a forgotten culture dish, contaminated by a mould — the most ordinary accident in the trade, and the gesture that follows is a reflex: the sink. Thousands of researchers had seen this scene before him. Every one of them rinsed the dish.
Fleming looks. Around the mould, a halo: the bacteria are no longer growing. He has just seen penicillin.
We tell this story as a charming exception. That is a misreading. Open a medicine cabinet: a considerable share of what it holds was born of an accident seen by the right person.
Lithium, the mainstay of bipolar disorder for three quarters of a century: John Cade, in 1949, was hunting a toxic factor in the urine of manic patients ; uric acid would not dissolve, so he took its most soluble salt — the lithium one. It was the unexpected lethargy of his guinea pigs that stopped him. Cisplatin, one of the most prescribed chemotherapies in the world: Barnett Rosenberg was not studying cancer, he was studying the effect of electric fields on bacteria ; his platinum electrodes were dissolving without his knowledge, and the bacteria stopped dividing. Warfarin, a major anticoagulant: cattle bleeding to death after eating spoiled sweet clover — a livestock disease described in the 1920s, whose active principle a Wisconsin chemist isolated the following decade. Chlorpromazine, which emptied the asylums: a French naval surgeon, Henri Laborit, was using it to prevent surgical shock and noticed a strange detachment in his patients. The first benzodiazepine: a compound synthesised then abandoned in 1955, found two years later at the back of a cupboard during a laboratory clear-out ; an assistant asked whether to throw it away, and Leo Sternbach said to send it for testing. Sildenafil, finally — Viagra: a molecule designed against angina, with weak clinical results, saved by an adverse effect volunteers kept reporting — erections.
Not one of these molecules was the object of the research that found it. All were seen by someone who was looking for something else.
What 2026 is optimising
Against this backdrop, the state of the art is impressive, and it deserves to be said without irony. Artificial intelligence has entered the core of the pharmaceutical reactor: target identification now happens in silico before any bench validation, platforms cross genomic, proteomic and transcriptomic data to reveal mechanisms that stayed hidden when those datasets were analysed separately, and molecules designed by AI are reaching late-stage trials ; 2026 will be the year Phase III results say whether the promise holds at scale. The sector has moved, the reports write, from the era of pilots to the era of builders — closed-loop discovery systems in which model and laboratory feed each other continuously.
But read those same reports to the end, and one sentence stops you. In its January 2026 outlook, Drug Discovery News concludes that discovery is becoming “more predictive, more integrated, and less tolerant of open-ended experimentation”.
Less tolerant of open-ended experimentation. It is written as progress, and it is — for everything we are looking for. The paradox is that we have just read the history of pharmacy: its greatest victories are precisely what nobody was looking for. A system trained to predict converges on the probable ; an optimised pipeline eliminates the outlier, the off-topic, the forgotten dish — everything that, for a century, has supplied the revolutions. We are building machines to find what we are looking for, of unprecedented power. Nobody is building the machine that sees what we were not looking for.
The problem reaches well beyond pharmacy, and it has a simple formulation. Search is a rear-view mirror: if you know what to type in the box, you have already done half the journey. The danger is not what you search badly — it is what you do not know you do not know.
What chance actually requires
Pasteur gave the formula in 1854: in the fields of observation, chance favours only the prepared mind. The phrase is always quoted ; we forget that it is a technical specification. Read each accident on the list again: none is pure luck. Each is the meeting of three conditions.
First, a collision — two worlds touching when nothing destined them to: a mould and a staphylococcus culture, platinum electrodes and cell division, a naval anaesthetic and psychiatry. Great discoveries are transverse or they are not ; they are born at the borders between disciplines, exactly where organisations file silos.
Second, a ground of held truth. Fleming knew exactly what should grow in that dish, and how — the anomaly was legible only against that ground of certainties. The same mould, in a poorly kept laboratory, would have been nothing but mould. Cade knew the normal behaviour of his guinea pigs ; Rosenberg, the normal growth of his bacteria. An anomaly exists only relative to a reliable expectation: in a world where anything may be false, nothing stands out any more.
Third, a memory that connects. Sternbach's compound was not discovered : it was found again, and someone judged that it deserved one more test rather than the bin. The Wisconsin vets connected cattle haemorrhages to a fodder because someone was holding the thread. Serendipity is not the accident: it is the recognition of the accident, and recognition is an act of memory.
Collision, ground, memory. Remove one of the three and the accident becomes noise again — rinsed down the sink with the dish.
The break is not faster prediction
This is why the next decade of discovery will not be played out where everyone is investing. Prediction is industrialising fast, and so much the better: it compresses the known. But the unforeseen — the whole class of penicillins, cisplatins and sildenafils still waiting — cannot be predicted, by definition. It is cultivated. And the three conditions that cultivate it are, very precisely, infrastructure problems.
Organising collisions. Continuously crossing an organisation's own corpus with the knowledge of the world: six hundred million scientific publications are worth nothing stacked ; they are worth something when a pharmacovigilance signal meets a materials-chemistry paper that nobody, in any silo, would have opened. The mechanism is not mysterious: atoms of knowledge do not assemble at random, they attract each other by affinity of structure. A maintenance report in the North Sea can resonate at the same frequency as a trading note in London — no human would have made the link, no search engine would have surfaced it, and the cognitive structure was the same nonetheless.
Holding the ground. Every claim verified, sourced, dated, replayable years later — so that the outlier stands out from the noise. Anomalies are visible only against a reliable ground, and that requirement is not a compliance constraint: it is the physical condition of discovery.
Preserving the memory of reasoning. Abandoned hypotheses, side effects noted in the margin, intuitions never pursued, trials stopped. Sternbach's compound slept two years in a cupboard and nearly went into the bin ; how many sleep today in reports nobody reads again, with nobody to hold the thread?
None of these three layers is a better model. They are the conditions under which any model — and any human — can see what they were not looking for. The industry is optimising the telescope. The seam is in the observatory: the kept place, where notebooks answer one another, where the outlier leaps out and the invisible finally becomes legible.
Chance cannot be industrialised. The prepared mind can be — it is, in fact, the only thing Pasteur left us to build.
This is the wager of (Urs): an infrastructure where knowledge binds by affinity across silos, on a ground where every fact is verified and every line of reasoning preserved — so that accidental discovery stops being an accident. Prediction compresses the known ; serendipity, for its part, can be equipped. Explore the infrastructure →
Sources. Fleming, British Journal of Experimental Pathology, 1929. Cade, “Lithium salts in the treatment of psychotic excitement”, Medical Journal of Australia, 1949. Rosenberg, Van Camp & Krigas, “Inhibition of cell division in Escherichia coli by electrolysis products from a platinum electrode”, Nature 205, 1965 ; antitumour effect demonstrated in 1969. Warfarin: sweet-clover haemorrhagic disease described in the 1920s, active principle isolated by K. P. Link at the University of Wisconsin in the 1930s. Laborit, on the use of chlorpromazine in anaesthesia and its psychiatric interest, 1952. Sternbach: compound Ro 5-0690 synthesised then abandoned in 1955, found again in April 1957 during a laboratory clear-out by his colleague Earl Reeder, marketed as Librium in 1960. Sildenafil: Pfizer clinical trials from 1991, redirected to erectile dysfunction after adverse-effect reports, FDA approval March 1998 (Ghofrani, Osterloh & Grimminger, Nature Reviews Drug Discovery, 2006). Pasteur, Lille address, 7 December 1854. State of the art 2026: Drug Target Review, “2026: the year AI stops being optional in drug discovery” (April 2026) and “AI in drug discovery: predictions for 2026” ; Drug Discovery News, “New tools and tougher economics will define drug discovery in 2026” (January 2026), the source of the quoted formula, and “The 2026 AI power shift” (Benchling report, “builder” phase).